Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-2x4c8-minmax-sse2-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <emmintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_2x4c8__sse2_ld128(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 2);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 float* c0 = c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr != 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51
52 const __m128 vmin = _mm_set1_ps(params->scalar.min);
53 const __m128 vmax = _mm_set1_ps(params->scalar.max);
54 XNN_FORCE_REALIZATION(vmin);
55 XNN_FORCE_REALIZATION(vmax);
56
57 const __m128i vmask = _mm_set1_epi8(0xF0);
58 XNN_FORCE_REALIZATION(vmask);
59
60 do {
61 const __m128i vksum = _mm_load_si128((const __m128i*) w);
62 const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
63 const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
64 const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));
65 const __m128i vzero = _mm_setzero_si128();
66 const __m128i vksum_lo = _mm_srli_epi32(_mm_slli_epi32(vksum, 16), 16);
67 const __m128i vksum_hi = _mm_srli_epi32(vksum, 16);
68
69 __m128i vzpprodksumhi0 = _mm_mulhi_epu16(vinput_zero_point0, vksum_lo);
70 const __m128i vzpprodksumlo0 = _mm_mullo_epi16(vinput_zero_point0, vksum_lo);
71 __m128i vzpprodksumhi1 = _mm_mulhi_epu16(vinput_zero_point1, vksum_lo);
72 const __m128i vzpprodksumlo1 = _mm_mullo_epi16(vinput_zero_point1, vksum_lo);
73
74 vzpprodksumhi0 = _mm_add_epi16(vzpprodksumhi0, _mm_mullo_epi16(vinput_zero_point0, vksum_hi));
75 vzpprodksumhi0 = _mm_sub_epi16(vzpprodksumhi0, _mm_and_si128(_mm_srai_epi16(vinput_zero_point0, 15), vksum_lo));
76 vzpprodksumhi1 = _mm_add_epi16(vzpprodksumhi1, _mm_mullo_epi16(vinput_zero_point1, vksum_hi));
77 vzpprodksumhi1 = _mm_sub_epi16(vzpprodksumhi1, _mm_and_si128(_mm_srai_epi16(vinput_zero_point1, 15), vksum_lo));
78
79 vzpprodksumhi0 = _mm_slli_si128(vzpprodksumhi0, 2);
80 vzpprodksumhi1 = _mm_slli_si128(vzpprodksumhi1, 2);
81
82 const __m128i vksumzp0 = _mm_or_si128(vzpprodksumhi0, vzpprodksumlo0);
83 const __m128i vksumzp1 = _mm_or_si128(vzpprodksumhi1, vzpprodksumlo1);
84
85 const __m128i vksum010 = _mm_unpacklo_epi32(vksumzp0, vzero);
86 const __m128i vksum230 = _mm_unpackhi_epi32(vksumzp0, vzero);
87 const __m128i vksum011 = _mm_unpacklo_epi32(vksumzp1, vzero);
88 const __m128i vksum231 = _mm_unpackhi_epi32(vksumzp1, vzero);
89
90 __m128i vacc0x0 = _mm_unpacklo_epi64(vksum010, vzero);
91 __m128i vacc0x1 = _mm_unpackhi_epi64(vksum010, vzero);
92 __m128i vacc1x0 = _mm_unpacklo_epi64(vksum011, vzero);
93 __m128i vacc1x1 = _mm_unpackhi_epi64(vksum011, vzero);
94
95 __m128i vacc0x2 = _mm_unpacklo_epi64(vksum230, vzero);
96 __m128i vacc0x3 = _mm_unpackhi_epi64(vksum230, vzero);
97 __m128i vacc1x2 = _mm_unpacklo_epi64(vksum231, vzero);
98 __m128i vacc1x3 = _mm_unpackhi_epi64(vksum231, vzero);
99 w = (const int32_t*) w + 4;
100
101 size_t k = kc;
102
103 while (k >= 16 * sizeof(int8_t)) {
104 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
105 const __m128i vxa0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c0, va0c0), 8);
106 a0 += 8;
107 const __m128i va1c0 = _mm_loadl_epi64((const __m128i*) a1);
108 const __m128i vxa1c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va1c0, va1c0), 8);
109 a1 += 8;
110
111 const __m128i vb01c01 = _mm_load_si128((const __m128i*) w);
112 const __m128i vbs01c0 = _mm_slli_epi32(vb01c01, 4);
113 const __m128i vb01c0 = _mm_and_si128(vbs01c0, vmask);
114 const __m128i vsb01c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c0);
115 const __m128i vxb0c0 = _mm_unpacklo_epi8(vb01c0, vsb01c0);
116 const __m128i vxb1c0 = _mm_unpackhi_epi8(vb01c0, vsb01c0);
117
118 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
119 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
120 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c0, vxb0c0));
121 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c0, vxb1c0));
122 const __m128i vb23c01 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
123 const __m128i vbs23c0 = _mm_slli_epi32(vb23c01, 4);
124 const __m128i vb23c0 = _mm_and_si128(vbs23c0, vmask);
125 const __m128i vsb23c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c0);
126 const __m128i vxb2c0 = _mm_unpacklo_epi8(vb23c0, vsb23c0);
127 const __m128i vxb3c0 = _mm_unpackhi_epi8(vb23c0, vsb23c0);
128
129 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
130 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
131 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c0, vxb2c0));
132 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c0, vxb3c0));
133
134 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
135 const __m128i vxa0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c1, va0c1), 8);
136 a0 += 8;
137 const __m128i va1c1 = _mm_loadl_epi64((const __m128i*) a1);
138 const __m128i vxa1c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1c1, va1c1), 8);
139 a1 += 8;
140
141 const __m128i vb01c1 = _mm_and_si128(vb01c01, vmask);
142 const __m128i vsb01c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c1);
143 const __m128i vxb0c1 = _mm_unpacklo_epi8(vb01c1, vsb01c1);
144 const __m128i vxb1c1 = _mm_unpackhi_epi8(vb01c1, vsb01c1);
145
146 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
147 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
148 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1c1, vxb0c1));
149 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1c1, vxb1c1));
150 const __m128i vb23c1 = _mm_and_si128(vb23c01, vmask);
151 const __m128i vsb23c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c1);
152 const __m128i vxb2c1 = _mm_unpacklo_epi8(vb23c1, vsb23c1);
153 const __m128i vxb3c1 = _mm_unpackhi_epi8(vb23c1, vsb23c1);
154
155 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
156 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
157 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1c1, vxb2c1));
158 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1c1, vxb3c1));
159
160
161 w = (const int8_t*) w + 32;
162 k -= 16 * sizeof(int8_t);
163 }
164
165 while (k >= 8 * sizeof(int8_t)) {
166 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
167 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
168 a0 += 8;
169 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
170 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
171 a1 += 8;
172
173 __m128i vb01 = _mm_load_si128((const __m128i*) w);
174 vb01 = _mm_slli_epi32(vb01, 4);
175 vb01 = _mm_and_si128(vb01, vmask);
176
177 const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
178 const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
179 const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
180
181 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
182 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
183 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
184 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
185 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
186 vb23 = _mm_slli_epi32(vb23, 4);
187 vb23 = _mm_and_si128(vb23, vmask);
188
189 const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
190 const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
191 const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
192
193 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
194 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
195 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
196 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
197
198 w = (const int8_t*) w + 32;
199 k -= 8 * sizeof(int8_t);
200 }
201
202 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
203 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
204 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
205 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
206
207 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
208 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
209
210 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
211 vacc1x0123 = _mm_srai_epi32(vacc1x0123, 4);
212 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
213 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
214
215 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
216 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
217 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
218
219 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
220 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
221
222 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
223 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
224 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
225
226 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
227 w = (const float*) w + 8;
228 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
229 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
230
231 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
232 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
233
234 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
235 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
236
237 if XNN_LIKELY(nc >= 4) {
238 _mm_storeu_ps(c0, vout0x0123);
239 _mm_storeu_ps(c1, vout1x0123);
240
241 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
242 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
243
244 c0 = (float*) ((uintptr_t) c0 + cn_stride);
245 c1 = (float*) ((uintptr_t) c1 + cn_stride);
246
247 nc -= 4;
248 } else {
249 if (nc & 2) {
250 _mm_storel_pi((__m64*) c0, vout0x0123);
251 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
252 c0 += 2;
253 _mm_storel_pi((__m64*) c1, vout1x0123);
254 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
255 c1 += 2;
256 }
257 if (nc & 1) {
258 _mm_store_ss(c0, vout0x0123);
259 _mm_store_ss(c1, vout1x0123);
260 }
261 nc = 0;
262 }
263 } while (nc != 0);
264 }